Enhancing MLP networks using a distributed data representation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Multilayer perceptron (MLP) networks often exhibit slow convergence due to the global nature of backpropagation and the inefficiency of hyperplane separation for nonlinear problems.
• Traditional MLP architectures struggle with complex, nonlinear decision boundaries, leading to prolonged training durations.

Purpose of the Study:

• To introduce a novel data representational approach, ensemble encoding, to address the slow convergence and inefficiency issues in MLP networks.
• To enhance the learning process by promoting local learning and improving nonlinear separability within the backpropagation framework.

Main Methods:

• Investigated the use of receptors with overlapping receptive fields as a preprocessing technique for encoding MLP network inputs.
• Implemented and evaluated the proposed ensemble encoding scheme on benchmark classification and time-series prediction problems.

Main Results:

• Ensemble encoding demonstrated the promotion of local learning and provided enhanced nonlinear separability for MLP networks.
• Simulations indicated a significant reduction in the training time required for MLP networks when using ensemble encoding.

Conclusions:

• Ensemble encoding is an effective data representational strategy that improves MLP network training efficiency and performance on nonlinear tasks.
• This nonlinear preprocessing technique is independent of the learning algorithm and MLP model architecture, offering a versatile alternative for various MLP applications.